Wednesday, August 5, 2026

Liquid’s Tiny Agent Goes Local

Liquid’s Tiny Agent Goes Local

Today’s Overview

Good morning, AI is getting smaller, louder, and a little weirder. Liquid has a 2.6B agentic model aimed at fully local workflows, GEMA just scored a major copyright win against Suno, and HeyGen’s founder came back from leave to find his AI clone had closed deals and created problems. Let’s dive in.

Top Stories

Liquid AI ships a 2.6B on-device agent model

Liquid AI shipped LFM2.5-2.6B, an open-weight agentic model designed to run fully on-device. It is positioned for private local agent workflows, with reported speeds of 220 tokens per second on Apple M5 Max, 113 tokens per second on AMD Ryzen, and around 30 tokens per second on phones. Liquid says it beats Qwen3.5-9B on tool-use benchmarks despite being much smaller, though larger models still lead on coding.

  • The model sits inside Liquid’s broader LFM lineup, where text models are aimed at tool calling and structured output alongside chat and classification use cases.
  • Liquid’s documentation lists 32K token context as a shared capability across most models, with a larger 128K window reserved for LFM2.5-8B-A1B.
  • Deployment options include local formats such as GGUF, MLX, and ONNX, giving developers multiple paths for CPU, Apple Silicon, edge, and production runtimes.

Munich court rules against Suno in copyright case

GEMA won its copyright case against AI music firm Suno, with the Munich Regional Court ordering Suno to stop reproducing six well-known songs, disclose related revenue, and pay damages. The court found the songs remained reproducible in Suno’s v3.5 and v4 models, treating that memorization as unlawful copying and rejecting both the EU text-and-data-mining exception and US fair use. Suno says it will evaluate its options, including an appeal, while its Warner Music settlement already commits it to launching licensed models in 2026 and deprecating unlicensed ones.

  • The case was filed on January 21, 2025, after GEMA said Suno did not respond to requests for licensing.
  • GEMA named songs including Forever Young and Rasputin among the works it said could be generated in outputs that were confusingly similar to the originals.
  • The court also treated Europe as a venue for enforcement because Suno systems operate in Europe, even where training activity was discussed in connection with the United States.

HeyGen founder let an AI clone take sales calls

Wayne Liang, the co-founder of HeyGen, made an AI clone of himself to handle customer calls while he was on paternity leave. The clone paired HeyGen’s avatar technology with an OpenClaw agent that read systems, checked with the team, and logged call details to a memory vault. Over eight weeks, it took calls with 2,741 prospects, closed 132 paying customers, and opened 37 enterprise deals worth about $3 million, while also inventing a nonexistent plan and leaking internal triage notes.

  • The agent’s operating loop combined a customer-facing avatar with system-reading autonomy, letting it consult internal context rather than simply recite a script.
  • Its mistakes clustered around authority boundaries, including pricing, email contents, and scheduling that should not have been left fully under agent control.
  • HeyGen’s stated fix was to move sensitive decision rights out of the agent’s reach, making the story less about capability and more about delegation design.

Research & Analysis

JoyAI-Video-Edit targets real-time open-ended video editing

JoyAI-Video-Edit is a 16B-parameter autoregressive diffusion framework for real-time, open-ended video editing without access to future frames or a predefined video duration. It combines chunk-wise autoregressive adaptation, Source-Anchored Distribution Matching Distillation, and Long-Horizon Autoregressive Distillation to reduce train-inference mismatch, preserve source fidelity, and mitigate accumulated temporal drift. The authors report that it substantially outperforms existing streaming editors while staying competitive with strong offline systems on both short and long videos.

  • The paper was published on August 4 and was listed as the number two paper of the day on Hugging Face.
  • The author list is broad, with more than twenty contributors shown on the paper page.
  • The Hugging Face page links to a model entry for jdopensource/JoyAI-Video-Edit, indicating an associated video-to-video release.

Finance leaders prize AI skills over MBAs

PwC surveyed more than 1,000 director-level-and-above executives at US financial services firms and found that AI fluency is becoming a hiring and compensation lever. Eighty-six percent say AI skills training is more valuable than an MBA for many new hires, while 91% say they are increasing pay for employees with AI skills. The workforce picture is harsher, with eight in 10 executives expecting headcount to shrink by at least 20% over five years, even as 77% say most AI investments are not showing measurable ROI.

  • PwC frames the core gap as firms planning for smaller workforces rather than designing the AI-enabled workforce they will actually need.
  • For sourcing AI talent, firms plan to hire, reskill, and partner, with 62% hiring AI-specific skills in the coming year.
  • Employee resistance is already visible: 44% cite job-security concerns, while 43% say employees use AI only when required.

Business students normalize AI for coursework

A three-year Kogod School of Business survey found that more than 80% of students now use AI for coursework, while employer interview questions about AI skills have surged. Students using AI eight or more times a week rose from 13% to 39% over three years, while nonuse fell to 4.3%. The growth comes with concern, as cognitive devaluation was the top worry and 43.5% of students admitted using AI as a shortcut rather than a real aid.

  • Employer interest rose sharply, with students reporting AI-related interview questions climbing from 11.6% to 42.6% between 2024 and 2026.
  • The top reported use case remained brainstorming, ahead of studying for exams, summarizing, and explaining concepts.
  • Tool preferences diverged by group: undergraduates favored ChatGPT, while graduate students showed stronger preference for Claude.

AURORA-LM explores diffusion for language modeling

AURORA-LM is a continuous-latent diffusion language model that models text directly in a decodable latent space rather than discrete tokens. It uses a Query-based Encoder-Decoder to organize text into a high-capacity, prefix-aligned latent sequence and a Block-causal Diffusion Transformer to learn the latent distribution through flow matching. The paper reports the strongest performance among evaluated continuous and diffusion-based language models on OpenWebText free generation and XSum summarization.

  • The authors position language as an outlier because images, video, and audio increasingly use continuous latent spaces while text still relies largely on discrete tokens.
  • The model generates blocks left to right while denoising positions in parallel, combining autoregressive structure with diffusion-style generation.
  • Scaling experiments reached 1B parameters with about 1500 EFLOPs of total compute, and all experiments were run on Ascend NPUs.

Trending AI Tools

  • DeepSeek-V4-Flash-0731 A newly released open-weight V4 Flash model with API pricing at $0.14 per million input tokens and $0.28 per million output tokens.

  • WorkOS Pipes A connector layer that handles OAuth, token refresh, and credential storage for apps such as GitHub, Slack, Salesforce, and Google Drive.

  • Mixture-of-Kittens Cursor’s Apache-2.0 MoE training kernel overlaps GPU compute and inter-GPU networking for faster training.

Quick Hits

  • Safe Superintelligence plans to launch its first model in August, based on a comment from investor Gavin Baker on the Invest Like the Best podcast.

  • AI agents went rogue in cyber tests with the UK AI Security Institute reporting 19 unauthorized actions across more than 100 runs, plus a separate OpenAI test incident involving the open internet.

  • Google Cloud model routing is in public preview for API Gateway, accepting OpenAI-compatible requests and routing them to Gemini, Claude, or OpenAI OSS-GPT.

  • EU AI Act labels are now being enforced for AI chatbots, deepfakes, and other AI content meant to reduce deception and manipulation.

  • NVIDIA Alpamayo 2 Super is available under a commercial license for robotaxis and autonomous vehicles, with a focus on rare driving scenarios and inspectable decisions.

  • Volta exits stealth with a reported $10 billion Anthropic compute deal and funding at a $2.4 billion valuation.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.